nshconfig

nshconfig is a small lifecycle layer over Pydantic for typed, Python-first ML configuration. Pydantic defines and validates the schema, and nshconfig re-exports its authoring API for a single import. nshconfig adds explicit mutable drafts, unbound parent-dependent templates, and declaration-ordered Python interpolation.

import nshconfig as C


class Optimizer(C.Config):
    learning_rate: float = 3e-4


class Run(C.Config):
    optimizer: Optimizer
    epochs: int


work = Run.config_draft()
work.optimizer.learning_rate = 1e-4
work.epochs = 100
run = work.config_finalize()

Calling Run(...) directly first performs ordinary Pydantic validation. A narrow missing-parent interpolation failure produces an inert template that binds when used in a concrete parent field; otherwise construction returns a final or raises the original validation error. Drafts are explicit, mutable, and may be incomplete. Finalization returns a fresh, field-frozen Pydantic model without consuming the draft.

There is no configuration language, registry, loader, provenance subsystem, record format, or code-generation layer. The semantic design is the authority for lifecycle, ordering, and structural graph behavior.